Tube MPC for Robust Constrained Control

• Tube Model Predictive Control is a robust control strategy that separates a deterministic nominal trajectory from error corrections using invariant tubes to manage bounded disturbances.
• The article details various tube parameterizations—including rigid, homothetic, polytopic, and zonotopic formulations—that balance computational efficiency and conservatism.
• It also explores system-level and learning-based extensions that ensure recursive feasibility, robust stability, and adaptive performance under model uncertainties.

Tube Model Predictive Control (MPC) is a foundational paradigm in robust constrained control, particularly for systems subject to bounded disturbances, model uncertainties, or both. The tube-based approach separates the tracking of a nominal (planned) system trajectory from the robust compensation of deviations (errors), leveraging invariant set theory and ancillary feedback control laws to guarantee constraint satisfaction and stability, even in the presence of significant uncertainty. This article systematically details the theory, methodologies, computational formulations, and major variants of tube-based MPC, emphasizing computational advances, scalable set parameterizations, and learning-based and adaptive extensions for both linear and nonlinear systems.

1. Fundamental Concepts and Problem Setting

Tube-based MPC frameworks are engineered to achieve robust satisfaction of state and input constraints in uncertain and/or disturbed dynamical systems by constraining the evolution of the closed-loop state within a sequence of time-varying sets, termed "tubes". The generic setting is a discrete-time dynamical system: $x_{k+1} = f(x_k, u_k, w_k)$ where $x_k$ is the state, $u_k$ the control, and $w_k$ a disturbance, process noise, or exogenous model uncertainty (possibly with complex, dynamic, or set-valued structure) (Tranos et al., 2022, Saccani et al., 2023, Das et al., 2024, Schwenkel et al., 2019, Sieber et al., 2021). Tubes are constructed so that for any realization of $w_k$ within an assumed bounded set, the trajectory remains feasible and within prescribed constraints.

Core elements:

• Nominal Trajectory: Planned sequence $z_k$, computed deterministically using a model (linearized or nominal, neglecting disturbances).
• Error Dynamics: Actual error $e_k = x_k - z_k$ governed by ancillary feedback correcting deviations.
• Tube Cross-sections: At each time, a set $E_k$ such that $x_k \in z_k \oplus E_k$.

Robustness is ensured by (i) tightening state/input constraints for the nominal trajectory using the maximal deviations characterized by the tube, and (ii) designing an ancillary controller (e.g., $u_k = v_k + K(x_k - z_k)$) so that all closed-loop errors remain inside the tube for all allowed disturbances.

2. Tube Parameterizations and Construction

Classical tube-based MPC relies on designing a robust positively invariant (RPI) set for the error dynamics using a fixed affine (often linear) feedback law. Advances in the field have led to diverse parameterizations and optimization-based tube designs:

2.1 Rigid and Homothetic Tubes

• Rigid tubes: The error tube is a fixed RPI set $\mathcal{S}$ calculated offline for the worst-case disturbance; the actual state is constrained to $x_k \in z_k \oplus \mathcal{S}$ (Gao et al., 2023).
• Homothetic tubes: The error tube is a scaled version of a fixed shape: $x_k \in z_k \oplus \alpha_k \mathcal{S}_0$ with scaling parameter $\alpha_k$ optimized online, introducing flexibility while maintaining fixed tube shape (Gao et al., 6 May 2025, Saccani et al., 2023).

2.2 Polytopic and Zonotopic Tubes

• Elastic/Polytopic tubes: The error tube cross-section is a general polytope, with parameters (facet offsets or scalings) optimized online, so the tube can "shrink" or "expand" directionally to minimize conservatism (Parsi et al., 2022, Diaconescu et al., 24 Sep 2025).
• Zonotopic tubes: The tube cross-section is a zonotope; online scaling or full generator updates enable efficient reachability analysis with low computational cost and excellent scalability (Diaconescu et al., 24 Sep 2025, Ghiasi et al., 24 Dec 2025, Alcala et al., 2020).

2.3 Ellipsoidal Tubes

The cross-section is a (homothetically) scaled ellipsoid parameterized by a common shape matrix and scalar radius, yielding online SDPs (semi-definite programs) of linear complexity in system dimension, especially when using LMI-based synthesis (Parsi et al., 2022, Schwenkel et al., 2021).

2.4 Configuration-Constrained Polytopic Tubes

A recent advance allows the tube shape (e.g., a polytope) to be freely optimized online, subject to vertex-configuration constraints ensuring that the facet–vertex structure remains unchanged. This joint parameterization of facets and vertices enables the tube to adjust in shape and size, strictly reducing conservatism compared to rigid, homothetic, and even elastic tubes (Villanueva et al., 2022, Badalamenti et al., 2024).

3. System Level Synthesis and On-line Tube Optimization

Recent tube-MPC literature promotes online optimization of the tube parameterization rather than relying only on offline-computed, fixed-structure sets.

3.1 System Level Parameterization (SLP) and System Level Tube-MPC (SLTMPC)

• SLP: Describes closed-loop responses to disturbances using block-lower-triangular matrices that encode the error system's entire trajectory (Sieber et al., 2021). The system-level approach enables direct convex optimization over these response maps.
• SLTMPC: Embeds the choice of tube controller (or feedback gain) into the system-level variables, jointly optimizing the nominal trajectory and error feedback online as a single QP. This intermediate approach bridges classical tube-MPC (lowest computational effort, most conservative) and full disturbance-feedback MPC (least conservative, highest online burden) (Sieber et al., 2021, Sieber et al., 2024).

3.2 Online and Asynchronous Computation

Fully online computation of error tube cross-sections (possibly under time-varying model/data/uncertainties) is computationally intensive. A two-stage asynchronous process can separate:

• Primary, real-time loop: solves a QP for the nominal trajectory under fixed (stored) tightened constraints.
• Secondary, slower loop: periodically updates tube cross-section parameters and the corresponding constraint tightening by solving larger optimization problems (e.g., full SLTMPC) (Sieber et al., 2022, Sieber et al., 2024).

Recursive feasibility is preserved via convex-combination memory and careful update rules for switching tube parameters (Sieber et al., 2022).

4. Handling Uncertainty: Additive, Multiplicative, and Dynamic Disturbances

Tube-MPC naturally accommodates a wide range of uncertainty structures:

4.1 Additive and Parametric Uncertainty

• Classical tube-MPC and many modern variants assume disturbances and model errors are bounded in known polytopic sets, with the ancillary feedback gain and tube designed for the largest possible realization (Sieber et al., 2021, Parsi et al., 2022).
• Advanced schemes accommodate polytopic (convex-hull) model uncertainty, potentially optimizing over auxiliary disturbance sets to jointly cover both additive and parametric errors (Sieber et al., 2024, Zhong et al., 2022).

4.2 Dynamic and Unmodeled Uncertainty

Integral quadratic constraint (IQC) theory provides a framework to bound the effect of dynamic or unmodeled uncertainty using multiplier filters and quadratic storage functions, enabling tube construction based on Lyapunov arguments that guarantee exponential stability and robust constraint satisfaction (Schwenkel et al., 2021).

4.3 Data-driven and Adaptive Tubes

When the true disturbance set is unknown, online learning algorithms can iteratively update (shrink) the disturbance set and the associated tube cross-section using observed disturbance samples, with statistical guarantees (probabilistic recursive feasibility via scenario theory) (Gao et al., 2023, Gao et al., 6 May 2025). Adaptive tube-MPC can embed online parameter identification (e.g., with least-squares or set-membership methods) to tighten tubes and ancillary feedback as uncertainty decreases (Tranos et al., 2022, Morozov et al., 2020, Ghiasi et al., 24 Dec 2025).

5. Theoretical Guarantees: Feasibility, Stability, and Performance

Tube MPC's reliability arises from rigorous set-invariance arguments and Lyapunov analysis, yielding several guarantees:

• Recursive Feasibility: If the tube-based optimization is feasible at initialization, all subsequent applications of the control law using updated tubes (or their convex combinations) preserve feasibility under disturbance and uncertainty—even as the underlying system or reference changes (Sieber et al., 2022, Parsi et al., 2022, Sieber et al., 2021).
• Robust Constraint Satisfaction: By design, the tightening of state and input constraints according to the error tube's maximal expected deviation ensures all hard constraints are satisfied for every disturbance realization admissible under the tube construction (Tranos et al., 2022, Luo et al., 2021, Saccani et al., 2023, Villanueva et al., 2022).
• Robust/Practical Stability: Under mild stabilizability and contractivity conditions, tube-MPC guarantees closed-loop practical stability, with asymptotic or input-to-state stability (ISS) properties as the tube contracts or remains invariant (including in the presence of parametric and dynamic uncertainty) (Wild et al., 2021, Villanueva et al., 2022, Sieber et al., 2024).
• Performance Bounds: Advances include turnpike arguments, demonstrating that robust economic tube-MPC without terminal costs still achieves near-steady-state performance under strict dissipativity and closed-loop reachability assumptions (Schwenkel et al., 2019).

6. Computational Methods, Scalability, and Trade-offs

Efficient representations and scalable optimization formulations are central:

6.1 Polyhedral/Zonotopic/Ellipsoidal Set Encoding

• Polyhedral tubes yield favorable tightness but can grow combinatorially in description complexity with state dimension.
• Zonotopic tubes, parameterized by a moderate number of generators and scales, admit linear-in-dimension scaling and very fast set-propagation algorithms.
• Ellipsoidal tubes yield convex LMIs with linear dependence on dimension and horizon, especially suited to high-dimensional and highly-uncertain systems (Parsi et al., 2022, Diaconescu et al., 24 Sep 2025, Ghiasi et al., 24 Dec 2025).

6.2 Trade-offs

• Larger, more flexible tubes (e.g., via fully parameterized vertex/facet updates or online generator scaling) reduce conservatism, enlarging the domain of attraction and improving performance (Diaconescu et al., 24 Sep 2025, Villanueva et al., 2022, Badalamenti et al., 2024).
• However, this introduces additional online complexity (decision variables and constraints per step), balanced by techniques such as pre-computing support-function matrices, memory-based convex-combination, or reducing the parameterization's degrees of freedom.
• The choice between rigid/homothetic tubes, zonotopic, polytopic, and configuration-constrained formulations should be made according to dimensionality, computational resources, and robustness/performance requirements.

7. Extensions: Learning, Distributional Robustness, and Nonlinear Systems

• Learning-based and adaptive tube-MPC: Tightens the disturbance set and the error tube online with statistical or scenario-theoretic guarantees, recovering nominal performance as system knowledge improves while retaining robust feasibility (Gao et al., 2023, Gao et al., 6 May 2025, Tranos et al., 2022).
• Distributionally Robust Tube MPC: Handles unknown disturbance distributions using data-driven ambiguity sets (e.g., Wasserstein balls centered at the empirical law), leading to worst-case robust solutions over all plausible distributions, extending feasible, safe MPC to highly uncertain or data-driven environments (Zhong et al., 2022).
• Nonlinear and LPV Systems: Tube-MPC naturally extends to LPV/affine polytopic systems using scenario-based or heterogeneously parameterized tubes, allowing mixture policies that trade off performance and complexity (Hanema et al., 2019). Nonlinear systems are handled by sequential linearization, tube propagation under local feedbacks, and error tube scaling, retaining recursive feasibility and robust constraints (Schwenkel et al., 2019, Luo et al., 2021, Luo et al., 2024).

---

References (arXiv IDs provided according to data):

• (Schwenkel et al., 2019) Robust Economic Model Predictive Control without Terminal Conditions
• (Sieber et al., 2021) A System Level Approach to Tube-based Model Predictive Control
• (Sieber et al., 2024) Computationally Efficient System Level Tube-MPC for Uncertain Systems
• (Sieber et al., 2022) Asynchronous Computation of Tube-based Model Predictive Control
• (Parsi et al., 2022) Scalable tube model predictive control of uncertain linear systems using ellipsoidal sets
• (Diaconescu et al., 24 Sep 2025) Zonotope-Based Elastic Tube Model Predictive Control
• (Ghiasi et al., 24 Dec 2025) Safe Navigation with Zonotopic Tubes: An Elastic Tube-based MPC Framework
• (Morozov et al., 2020) Performance Analysis of Adaptive Dynamic Tube MPC
• (Villanueva et al., 2022) Configuration-Constrained Tube MPC
• (Badalamenti et al., 2024) Configuration-Constrained Tube MPC for Tracking
• (Tranos et al., 2022) Self-Tuning Tube-based Model Predictive Control
• (Gao et al., 2023) Learning-based Rigid Tube Model Predictive Control
• (Gao et al., 6 May 2025) Learning-based Homothetic Tube MPC
• (Schwenkel et al., 2021) Model predictive control for linear uncertain systems using integral quadratic constraints
• (Saccani et al., 2023) Homothetic tube model predictive control with multi-step predictors
• (Hanema et al., 2019) Heterogeneously parameterized tube model predictive control for LPV systems
• (Luo et al., 2021, Luo et al., 2024) Tube-based MPC for robotic manipulators
• (Zhong et al., 2022) Tube-based Distributionally Robust Model Predictive Control for Nonlinear Process Systems via Linearization